Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Multi-protocol consensus for agent swarms supporting Raft leader election, Byzantine fault tolerance, Gossip state propagation, and CRDT conflict-free merging.
.claude/skills/a5c-ai-consensus-mechanisms/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -48% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -21% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -52% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 17% | 0% |
| Protocol | Use Case | Fault Tolerance | Complexity | |----------|----------|-----------------|------------| | Raft | Leader-based consensus, ordered log | Crash faults (f < n/2) | Medium | | Byzantine | Untrusted agents, adversarial conditions | Byzantine faults (f < n/3) | High | | Gossip | Eventual consistency, state propagation | Partition tolerant | Low | | CRDT | Conflict-free replicated data types | Always convergent | Low |
agents/swarm-coordinator/ - Protocol orchestrationagents/strategic-queen/ - Weighted voting leadershipInvoke via babysitter process: methodologies/ruflo/ruflo-swarm-coordination
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 13,618 | 6,003 | -56% | 1 | 1 | 0% | 2,083 | 1,298 | -38% | 0 | 0 | — |
case-02 | fail→pass | 31,527 | 13,619 | -57% | 1 | 1 | 0% | 5,349 | 2,755 | -48% | 0 | 0 | — |
case-03 | fail→pass | 19,300 | 14,802 | -23% | 1 | 1 | 0% | 3,995 | 3,147 | -21% | 0 | 0 | — |
case-04 | fail→fail | 21,628 | 17,340 | -20% | 1 | 1 | 0% | 3,371 | 3,385 | +0% | 0 | 0 | — |
case-05 | fail→fail | 22,970 | 18,503 | -19% | 1 | 1 | 0% | 3,750 | 3,756 | +0% | 0 | 0 | — |
case-06 | fail→fail | 15,483 | 13,930 | -10% | 1 | 1 | 0% | 2,904 | 3,048 | +5% | 0 | 0 | — |
case-07 | pass→pass | 23,048 | 4,657 | -80% | 1 | 1 | 0% | 1,866 | 1,022 | -45% | 0 | 0 | — |
case-08 | fail→fail | 10,925 | 5,296 | -52% | 1 | 1 | 0% | 1,801 | 1,048 | -42% | 0 | 0 | — |
case-09 | fail→fail | 14,166 | 6,641 | -53% | 1 | 1 | 0% | 2,405 | 1,372 | -43% | 0 | 0 | — |
case-10 | fail→fail | 17,138 | 12,244 | -29% | 1 | 1 | 0% | 3,239 | 2,527 | -22% | 0 | 0 | — |
case-11 | fail→pass | 42,391 | 3,912 | -91% | 1 | 1 | 0% | 1,232 | 593 | -52% | 0 | 0 | — |
case-12 | fail→fail | 14,160 | 5,848 | -59% | 1 | 1 | 0% | 2,350 | 1,168 | -50% | 0 | 0 | — |
case-13 | fail→fail | 11,465 | 1,875 | -84% | 1 | 1 | 0% | 1,968 | 545 | -72% | 0 | 0 | — |
case-14 | fail→pass | 29,537 | 7,218 | -76% | 1 | 1 | 0% | 1,271 | 1,486 | +17% | 0 | 0 | — |
case-15 | fail→pass | 4,427 | 1,394 | -69% | 1 | 1 | 0% | 681 | 446 | -35% | 0 | 0 | — |
case-16 | fail→fail | 16,374 | 20,048 | +22% | 1 | 1 | 0% | 2,896 | 3,665 | +27% | 0 | 0 | — |
case-17 | fail→pass | 9,789 | 3,022 | -69% | 1 | 1 | 0% | 1,591 | 771 | -52% | 0 | 0 | — |
case-18 | fail→pass | 8,240 | 2,932 | -64% | 1 | 1 | 0% | 1,294 | 782 | -40% | 0 | 0 | — |
case-19 | fail→fail | 10,410 | 8,072 | -22% | 1 | 1 | 0% | 1,760 | 1,470 | -16% | 0 | 0 | — |
case-20 | fail→pass | 10,639 | 3,340 | -69% | 1 | 1 | 0% | 1,803 | 838 | -54% | 0 | 0 | — |
case-21 | fail→fail | 11,469 | 5,330 | -54% | 1 | 1 | 0% | 1,931 | 1,100 | -43% | 0 | 0 | — |
case-22 | fail→pass | 16,540 | 9,270 | -44% | 1 | 1 | 0% | 2,818 | 2,096 | -26% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +45 percentage points is the difference between those two pass rates over the 21 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.